摘要

This paper presents a novel Hough context model for detecting and localizing instances of a certain object class. In particular, our approach detects the object in parameter space by applying Hough voting in context. First, in Hough context, the features around object center or a point in parameter space are sampled and organized in context sense respectively for both training images and test images. Then a parametric discriminant function is constructed based on the Hough context. Such function scores the existence likelihood of object centers in parameter space. Finally, we formulate the training process of this discriminative model as a structure learning problem, which has already been well solved. Compared with the ISM-related methods that the voting is performed by the features independently, the core contribution of our method is that the voting based on context information is available. The experiments on several popular and challenging object datasets (i.e., ETHZ Shape, UIUC Cars, and PASCAL VOC 2007) demonstrate that the detection accuracy can be improved and the voting speed are impressively accelerated via using Hough context.